Hi everyone,
I’m using Qwen Image Edit with mask-based inpainting through diffusers (Python script on RunPod, not ComfyUI). My goal is object removal, for example removing a rug from a room photo and filling the area with the floor underneath.
What works well:
On smooth, plain surfaces like a painted wall, a plain ceiling, or a flat-colored floor, the inpainting is very good. The filled area blends in and you can’t tell anything was removed.
What doesn’t work:
The problems start when the area behind the object has texture, patterns, or lighting variation.
- Floor texture and design: On wooden planks, tiles, or marble, the inpainted area doesn’t continue the existing pattern. Plank lines don’t align, the grain or tile size is different, and the pattern direction sometimes changes.
- Light and shadows: The original floor has natural light falloff, reflections, and soft shadows. The inpainted area is either too flat and evenly lit, or it has lighting that doesn’t match the room. Sometimes the object’s shadow stays behind, or a strange new shadow appears.
- Color and sharpness: The filled area is often slightly different in brightness or color, and smoother or blurrier than the real photo. It looks “painted” even when the shape is right.
My setup:
- Model: [Qwen/Qwen-Image-Edit or Qwen-Image-Edit-2511]
- Pipeline: [QwenImageEditInpaintPipeline ]
- Steps: [e.g. 16], true_cfg_scale: [e.g. 4.0], strength: [e.g. 1.0]
- Prompt: [your prompt, e.g. “remove the rug and show the wooden floor underneath, same texture and lighting”]
How can i get better or real inpainting ? Please help !